A digital human company decided to build its own voice model. On October 9, HeyGen announced HeyGen Voice, its first in-house text-to-speech model, and it landed at the top of Artificial Analysis's Controlled Voice leaderboard. The official framing is not modest. CTO Rong Yan put it directly: "Debuting at #1 on Artificial Analysis is independent proof that you don't have to trade authenticity for quality."
The operative phrase is independent proof. The long-running argument in voice cloning has always been whether fidelity and pleasantness can coexist: clone the voice accurately and it sounds like a reading machine; make it pleasant to hear and the speaker's identity starts dissolving. HeyGen's claim is that its model delivers both, and that a leaderboard position is third-party evidence. Before asking what that first place means, though, the measurement conditions have to be pinned down exactly -- otherwise the number reads as marketing copy.
What the Leaderboard Result Actually Measures
The Artificial Analysis Controlled Voice leaderboard carries three conditions that belong in any summary of the result.
First, this is the Controlled Voice board, not an aggregate TTS ranking and not a language-coverage scorecard.
Second, the test language is English. Chinese, Spanish, and Japanese performance sit outside this particular score.
Third -- and this is the most informative condition -- the run compared the same eight cloned voices. That detail determines what the board is actually measuring. It is not asking which model's built-in default voice sounds nicest; it is asking how faithfully a model can reproduce a voice that does not belong to it. In other words, this is a leaderboard about clone preservation, not about stock timbre quality. For a company whose main business is digital humans, that is precisely the cell where it matters most and where scrutiny is warranted.
The models it leads are Qwen-Audio-3.1-TTS-Plus and Eleven v4 Turbo. Both names are worth pausing on: the former is a new flagship from the Chinese TTS camp, the latter a long-acknowledged leader in English speech. HeyGen sits ahead of both rather than tied with them.
The third-party reading deserves equal billing. This first place is an early score on the English controlled-voice board. It does not mean the model leads across every language, on latency, or in real-time conversation. The board answers one question -- how similarity holds up under English, with those eight cloned voices. It does not answer how Chinese cloning fares, what the first-packet latency is in milliseconds, or whether interruption and turn-taking feel natural. Reading a single-condition early result as an across-the-board victory compresses three separate gaps into one digit.
Why a Digital Human Company Built Its Own Voice
Avatar V is HeyGen's core model. Voice here is not an external dubbing service; it is a new format inside the same end-to-end stack as appearance, expression, and lip movement. Identity and voice come from the same source, and the direct benefit is synchronization: when timbre, mouth shapes, and expressions originate from three different models or datasets, misalignment is the normal case. A shared stack removes one whole layer of that problem.
The specific pain point on this pipeline has been described before in our digital human production walkthrough (AI Digital Human Creation SOP): lip sync that drifts, stiff expressions, a voice whose tone does not match the character's presence. The root cause usually is not editing -- it is that voice and visual identity are two separate things. Generating them inside one stack is the direct remedy. Compliance red lines for digital human output are covered in our industry look (AI Digital Human Wild West Era).
So in-house TTS carries two meanings for HeyGen. Externally, it takes voice quality back from the vendor's hands and removes the compromise between a third-party speech platform's output and its release schedule. Internally, it allows voice and avatar to be co-tuned, which is simply not achievable by bolting together two APIs. A side effect worth noting: an application company has started shipping models, meaning its control over the model layer has moved upward. That shift normally brings one good consequence and one bad -- faster iteration on its own schedule, but also absorbing inference costs, data-compliance responsibility, and the impossibility of blaming the upstream provider when something breaks. That trade-off is worth watching; it is not settled today.
The Decoupled Parameters: Clone Fidelity and Performance Intensity
This is the most substantive technical element of the release and also the easiest to miss.
At clone creation time, a parameter called similarity is set, determining how close the output tracks the original speaker. At synthesis time, a separate parameter called expressiveness_boost is passed, determining how intense the emotional delivery is. The two never appear at the same stage, and neither overrides the other.
The point of this split is that traditional TTS interfaces often bind similarity and expressiveness to a single knob. Turn it toward fidelity and the voice matches the speaker but flattens out; turn it toward performance and the emotion swells but the resemblance slips. Users face an irreversible trade-off between two goals, even though real requirements rarely look like that: a product read wants high similarity with moderate emotion, a narrative video wants slightly lower similarity with wider dynamic range. Decoupling converts one impossible choice into two independent adjustments.
The official recommended combination in HeyGen Studio is similarity set to max with expressiveness_boost at 0.5 -- a conservative starting point that restores the resemblance first and adds expression at low amplitude, rather than pushing emotion to maximum on the first take. The default also reveals a judgment: most users opening the product for the first time want "sounds like them," not "dramatic delivery."
Once decoupled, the number of control points increases as well. A single synthesis call used to expose one knob; now the same passage of text can sample one setting on the similarity axis and another on the expressiveness axis. One cloned voice can present differently across contexts while remaining recognizably the same voice -- the same set of vocal folds reading a notice and telling a story are two different jobs that no longer require two different voices.
Two Cloning Paths and the Parameter Error Boundary
HeyGen offers two cloning paths, and the difference is not just how much audio you record.
Instant voice cloning needs a single recording, uses the first three minutes of it, and completes in seconds. Its hard boundary is that it cannot be retrained, only rebuilt: once that voice exists, there is no way to add material and refine it in place. To improve the result, you start over with a new recording.
Professional voice cloning takes one to ten recordings totaling more than twenty minutes as training material, and it supports retraining. Retraining means you can add audio, iterate, and gradually bring a barely usable voice up to standard.
Both paths ultimately feed the same HeyGen Voice native TTS endpoint for synthesis -- same model producing the output, with the difference sitting in the upstream training stage and in long-term maintainability.
The detail most worth writing down is how the parameter system binds to clone type:
- Instant voices support expressiveness control;
- Professional voices support seed, speed, and pitch;
- Mismatching a parameter against the clone type returns a parameter error directly.
That is not a bug; it is capacity isolation by design. A professionally cloned voice has been through training and has a stable internal representation, which is what makes it safe to expose speed and pitch -- knobs that rewrite fundamental frequency and speaking rate outright. Instant cloning is a lightweight alignment, and emotional intensity is the only control it can honestly offer. The system does not guess at your intent: pass a parameter the voice does not possess and it errors out, rather than silently ignoring it or substituting an approximation.
For users, the actual value of this design is that mistakes surface before synthesis rather than in the finished cut. Pass pitch to an instant clone and the API returns a parameter error immediately, and you know what to change at once. Had the system silently degraded instead, you would receive audio at a strange tempo and spend an afternoon hunting for the misconfiguration. Integration teams should note this more carefully: instant and professional clones follow two different interface contracts, and mixing them fails on the first call rather than quietly producing wrong results on the third.
Pricing: Ninety-Nine Dollars a Month for the Professional Path
The model and API are open at no separate charge on the platform side -- HeyGen Voice is not billed independently and runs on platform quota. That is the official pricing position.
Professional voice cloning is an add-on service listed at $99 per month on the official site. What the fee covers is consent plus material: it requires the individual's explicit approval and uses between thirty minutes and three hours of that person's own voice as training data. The explicit consent here is not boilerplate. It is the compliance premise of the entire business -- what you are buying is not fidelity but a regular authorization chain. Whose voice may be cloned, and for what purpose, is handled at this layer as a question of responsibility rather than a technical parameter.
In industry context, professional cloning is a typical bundled price for material cost, training cost, and authorization liability. At $99 per month, what you get is licensed material, a formal authorization path, and the retraining capability described above. For someone who occasionally needs one voiceover, $99 per month sits above what instant cloning delivers for free; for a team maintaining a long-term commercial voice, the price falls in a reasonable band. Whether it pays depends on how long the voice needs to live -- the general framework covered in our voice cloning tool comparison, not a selection exercise this article repeats.
Company Numbers: All Self-Reported
The release materials include a set of company-level figures. To keep them from being misread, they are flagged here as self-reported by HeyGen and not independently verified: founded in 2020, more than 40 million users, 196 countries, over 118 million videos generated, adopted by 85% of the Fortune 100, and ARR above $200 million, doubled in eight months. Alongside these sit a G2 top ranking for most realistic in its category and Fast Company's 2026 most innovative companies designation.
These numbers establish HeyGen's position and commercial reach in digital humans, but they cannot be used to infer model quality. There is no verifiable conversion rate between user count, ARR, and a leaderboard position -- and how many of those 40 million users have touched the newly released in-house voice model has not been disclosed.
One widely cited industry statistic needs a source label as well: in a survey of more than 1,000 small business owners, 71.6% recorded videos they ultimately never published (reported by ai-bot, a single source with no independent corroboration located). That figure points to the gap between the cost of recording and the return from publishing -- a broad pain point in digital humans, not something caused by this model release.
How This Divides From Our Earlier Voice Coverage
We have already published two adjacent pieces on voice, and this one deliberately does not repeat them.
The AI voice cloning tools comparison puts six mainstream cloning tools side by side and answers which vendor to pick. The cloud versus local voice AI review works out the cost structure and data ownership question. This article is a release event plus a technical teardown, answering a third question: what HeyGen's technical choices themselves imply. For selection, read the first two; this piece offers no tool recommendations.
To go a layer deeper technically, the GPT-Live-1 voice API SOP covers hands-on integration for real-time speech interfaces, with attention to latency, interruption, and end-to-end timing. Notably, the HeyGen Voice announcement includes no latency figures on the real-time conversational side -- which makes that integration walkthrough and this release event a natural pair.
Closing
The technical interest in HeyGen Voice is not the first place. It is two concrete moves: separating similarity from expressiveness into independent parameters, and using parameter errors to draw the capacity boundary explicitly. The first addresses the long-standing claim that fidelity and pleasantness cannot both be had in voice cloning. The second expresses an engineering principle -- interfaces should not politely guess at user intent. Both say more about how this company builds models than a leaderboard position does.
As for what the number itself is worth, the answer depends on how you read what the board measures: English, eight cloned voices, similarity on a single dimension, at an early stage. It is enough to establish that HeyGen has entered the top tier for clone fidelity. It is not enough to claim leadership across languages, latency, and real-time conversation at once. Keeping those two sentences apart is the discipline any technical leaderboard reading requires.
Three things are worth tracking next: when controlled-voice results for Chinese and other languages become public; whether this decoupled two-parameter interface becomes an industry de facto standard, since the earlier a product binds to such a pattern, the cheaper its later migration; and whether the $99 per month professional authorization chain pulls industry pricing along with it. With the model inside the stack, the parameters decoupled, and the boundaries drawn by errors, what remains to compete over on the digital human chain is content quality itself.